World: r4wp

We are giving developers more control by saying that some stuff is 
under their control (i.e. the code blocks). We are providing some 
safety by saying that some stuff is presumed to be not under their 
control and thus possibly suspect (i.e. immediate-evaluation parameters 
to functions that they didn't write). We do screening of some stuff 
because that cuts down on the screening they have to do themselves. 
That way dvelopers can use functions and assume that they are safe 
to use by default.


For instance, one advantage of #1993 FOR would be that they would 
have to go out of their way to make it do an infinite loop, since 
no combination of start, end and bump would generate one. That means 
that they wouldn't have to wrap calls to FOR in expensive conditional 
code, they can just pass in any values of those parameters and trust 
FOR to never go infinite without them expecting it. Your CFOR would 
not have that advantage, but since it takes code blocks for all parameters 
it is assumed that you are more careful about those code blocks, 
as you should be as a general rule in R3.


It's about providing a balance. Complete consistency in how all parameters 
are treated regardless of their nature would not allow us to help 
developers where they need it. However, having a consistent policy 
that code must be treated more carefully by developers than non-code 
allows developers some flexibility while still allowing them to be 
careful. That is why code that developers provide explicitly is considered 
to be what they want to do, at least from the outside of functions. 
And you can make the distinction between code and non-code using 
simple type tests, which is why we have APPLY and ASSERT/type.

But that sill does not solve the case of finite loops, does it?

 - well, you could have really large finite loops that on a pragmatic 
 level are close enough to infinite to still cause a problem. I think 
 that there is a global evaluation limit that is supposed to protect 
 you from that, but I don't know if that currently works.

Of course the evaluation limit is under developer control.

THat is absoluteyl misunderstanding where the real problem is.

OK, nevermind. I will try to be understandable for you:


if you want to do so, you *can* eliminate the only combination of 
START, END, BUMP values that causes infinite loop by default while 
still remaining consistent at the cost of disallowing something that 
may be perceived as "legitimate" by some users, and as an "unsolicited 
for" limitation.

Really? Because I was assuming that FOR would have termination conditions 
(similar to yours). The initial conditions model just picks which 
set of termination conditions to apply, at least as far as the start, 
end and bump parameters are concerned. The body parameter, being 
a code block, is assumed to be under developer control. So if the 
developer wants to hack the termination conditions in the code block 
that is their fault.

Ha, are we still discussing 1993 or some other ticket?

FOREVER is assumed to be a solicited-for infinite loop, because it's 
right there in the name. #864 is assumed to be whatever the developer 
says, because "General loop" is right there in the doc string. #1993-1994 
FOR has the *feature* of not *accidentally* being infinite for any 
value of start, end and bump, its constraint is a feature; of course 
it could be *intentionally* infinite by changing the index in the 
code block, but that just means that there is one parameter that 
the developer would have to be careful about, the body block, and 
since that is a general pattern throughout R3 they would be doing 
that anyway.

Assuming that we are discussing #1993 and assuming:


* you insiste to allow the FOR I 1 1 0 and similar to be "infinite 
by default"

* want to support cycle variable changes in a simple and consistent 
way

the best you can do is to cause an error when finding out that

    all [start = end bump = 0]


is TRUE, since in that case there is no reasonable terminating condition 
that could not cause infinite loop by default. Some people may object, 
but otherwise it looks as not a big deal.

err: "you insist to not allow FOR I 1 1 0 and similar to be "infinite 
by default"" is what I meant

{you insiste to allow the FOR I 1 1 0 and similar to be "infinite 
by default"} -> {you insiste to allow the FOR I 1 1 0 and similar 
to never be "infinite by default"}

Right.

The fact that we even have a FOREVER loop at all means that there 
would be a value in having developers use it, just for their own 
documentation.

And that allows developers to be less careful with start, end and 
bump, for values of "careful" that mean "wrapping FOR in expensive 
conditional code".

We've been trying to reduce the amount of expensive conditional code 
throughout R3. That's why a lot more functions allow none now.

ASSERT/type means that typechecks don't have to be considered expensive 
anymore. Value checks still require EITHER and other conditional 
functions though.

For that matter, typesets make the type tests of function argument 
specs less expensive now too, so we can afford to be picky there 
as well.

Am I to understand that instead of causing an error int that case 
you simply prefer to not loop taking it as an exceptional condition?

An out-of-range condition, yes.

The other model was the trigger-an-error model.

BrianH: I'm really glad to hear that there is a concerted effort 
to reduce the amount of expensive conditional code throughout R3. 
 Great job all!

Hmm, it is not out-of-range in the normal sense, because when START 
= END and BUMP = 0 the values both are and remain "in range", but 
it is "exceptional"

Depends on the model. If bump is a velocity, you can say that only 
velocities above (or below of you're going in reverse) 0 are in range.

That is a matter of coming up with a plausible theoretical explanation 
for something that we want to do for practical reasons.

Then start-vs-end sets the direction, and bump sets the velocity. 
It's just a way to explain *why* to newbiees.

Dealing with the consequences of triggering an error is more expensive, 
so we tend to only want to trigger errors when they really *are* 
errors. If there is a plausible way to just do nothing and/or return 
none when it's not potentially damaging, we should come up with a 
rationale that lets do that instead.

You can say that you "support" zero velocity by "not looping", but, 
in fact, you rather don't support it by failing as silently as possible.

It's really a rationale.

No problem to say it is "zero velocity". The problem is that in "normal 
life" something having zero velocity does not "vanish", rather it 
stays where it is.

So you may be caught as using "inappropriate logic", but I do not 
mind, being able to eventually answer that it is an exceptional case 
that you simply did not want to handle to not cause headaches to 
some users, while causing inconveniences to people being able to 
calculate what it is they should have expected.

We can just arbitrarily declare that we want 0 velocity to be considered 
out of range, as a favor to the developer, and the velocity explanation 
gives us a good excuse to not trigger an error. FOREVER existing 
means that they have other options, and index setting means that 
they can do whatever they want if they really want to, so it's not 
actually a constraint if they don't want it to be.

You can always declare something arbitrarily. The problem is that 
if you do declare a + b = none in case a = 0 you are most probably 
causing inconveniences to all people knowing that there might have 
been a more consistent behaviour...

Remember, #864 is a proposal to replace FOR with a more flexible 
power-user function that would be less safe to use. They lose some 
safety as a tradeoff for more power and prettier sytnax. So, they 
lose two features (safety and backwards compatibility) but gain more 
flexibility. The greater flexibility would come at the expense of 
a slower function: negligably in the case of the function itself, 
but more when you add the conditional wrapper code, so it would have 
to be used carefully if you want it to be efficient. Overall, that 
is the R3 motto right there: R2 is for newbies, R3 for power users.

But let's just forget about it in this specific case. I guess that 
FOR is not expected to be used extensively anyway.

So, I would recommend that the #1993 restrictions against accidental 
infinite loops should go into R3/Backwards and rebol-patches, because 
R2 is for newbies. And I would recommend that #864 be the new FOR 
for R3 and R2/Forward, because R3 is for power users.

Man, you guys are typing faster than I'm reading. :-)

LOL

the best you can do is to cause an error when finding out that

 
   all [start = end bump = 0]

is TRUE, since in that case there 
is no reasonable terminating condition that could not cause infinite 
loop by default.


I think this answers what I was asking. Though it seems that Ladislav 
wants [1 1 0] to be infinite for consistency, while Biran and I want 
it not to be, for perceived user friendliness. :-)

Brian, for that case, does your model make it a "no loop" or "once 
only" condition?

Gregg, we just agreed to make it "no-loop"

OK, I skimmed too fast. Thanks.

(which is "the most silent failure")

once only

 would be rather inconsistent with any terminating condition, because 
 there is none that could cause it to loop exactly once

Updting %new-loop.r now.

Posted %mezz/new-loop.r

Gress, for the start-vs-end-sets-direction bump-is-velocity model:

* start=end means no direction so just loop until the =end termination 
condition is met and ignore bump. If the index gets changed in the 
body block, let the =end termination condition handle it.

* start<end means positive direction, for values of "positive" that 
don't include 0, so bump <= 0 is out of range, meaning no loop. The 
termination condition *if we start looping* is >= end.

* start>end means negative direction, for values of "negative" that 
don't include 0, so bump >= 0 is out of range, meaning no loop. The 
termination condition *if we start looping* is >= start.


Positive and negative directions don't include 0 because if the developer 
wanted to do an infinite loop they would have used FOREVER or R3's 
#864 FOR general loop. R2 was aimed at newbies, and they need extra 
coddling.

Gress -> Gregg

For the bump-sets-direction start-and-end-set-the-range model, 0 
doesn't set a direction so it should trigger an error. Otherwise, 
the same.

Brian, can you point out which test case is incorrect, and what it 
should produce? That way we can match against Ladislav's examples.

I like the model that doesn't trigger an error.